Verifiable Claim Ratio

The VCR is calculated as a simple fraction: VCR = (Number of Verifiable Claims) / (Total Number of Factual Claims). A claim is defined as a discrete, atomic statement of fact. The process involves:

• Claim Extraction: Using NLP to decompose generated text into individual factual assertions.
• Corpus Verification: Each claim is checked against a trusted corpus (e.g., a knowledge graph, vetted database, or primary source archive).
• Binary Classification: A claim is marked 'verifiable' only if direct, unambiguous supporting evidence is found. Unsupported or contradicted claims are not counted. A score of 0.95 means 95% of factual statements were confirmed.

VCR is a measure of groundability, not absolute truth. A claim can be verifiable (evidence exists) but still be factually incorrect if the source corpus contains errors. Conversely, a true claim is unverifiable if no evidence exists in the accessible corpus.

• Accuracy: Measures alignment with ground truth reality.
• VCR: Measures alignment with a specific, trusted evidence base. This distinction is crucial for systems using Retrieval-Augmented Generation (RAG), where the goal is fidelity to the provided context, not omniscience. A high VCR indicates strong citation integrity, not infallibility.

The VCR is entirely dependent on the trusted corpus used for verification. The same AI output can yield wildly different VCRs against different corpora.

• High-Authority Corpus: Using peer-reviewed journals and primary sources yields a rigorous, lower VCR that reflects academic credibility.
• Broad Web Corpus: Using a general search index may yield a higher VCR but risks validating claims against low-quality or unvetted sources. The choice of corpus defines the standard of evidence. A robust system must declare its verification corpus to make the VCR meaningful, linking the metric directly to a Source Tier Classification framework.

VCR is the inverse proxy for the Hallucination Rate for factual claims. If a model generates 100 factual claims and 8 cannot be verified, the VCR is 0.92, and the observed hallucination rate is 8%.

• Predictive Power: A consistently low VCR in testing is a strong predictor of poor Hallucination Risk Index scores in production.
• Non-Factual Hallucinations: VCR does not capture logical errors, incoherent reasoning, or stylistic hallucinations that don't manifest as discrete, checkable facts. VCR is a necessary but not sufficient metric for overall output quality, best used in conjunction with Confidence Calibration and Semantic Relevancy Vector analysis.

The atomicity of claim extraction dramatically impacts the VCR. A sentence like 'The Eiffel Tower, built in 1889, is in Paris' contains three claims:

1. The Eiffel Tower is in Paris.
2. The Eiffel Tower was built in 1889.
3. The structure is named 'The Eiffel Tower'. A sophisticated Attribution Granularity Level system will verify each independently. A naive approach might mark the whole sentence unverifiable if only one fact is unsupported, unfairly penalizing the score. High-fidelity VCR requires fine-grained, Claim-Source Alignment at the sub-sentence level.

Achieving a high VCR requires a systematic, multi-layered architecture:

• Knowledge Graph Grounding: Anchoring generation to a deterministic graph ensures claims are derived from curated facts, not statistical likelihood.
• Retrieval-Augmented Verification: A post-generation step that uses a separate, high-precision retriever to fact-check every claim against a Primary Source Priority index.
• Citation Chaining Protocol: Automatically tracing a generated claim back through its evidence chain to the original source, validating Evidence Chain Integrity. This transforms the system from a 'generator that sometimes cites' to a 'synthesizer that is strictly bound by evidence'.

A composite metric that quantifies the degree of semantic and factual correspondence between a specific AI-generated statement and the content of its cited source. This score directly feeds into the Verifiable Claim Ratio by determining whether a citation genuinely supports the claim it accompanies.

• Natural Language Inference (NLI) models detect entailment, contradiction, or neutrality
• Semantic similarity is measured via cosine similarity of dense vector embeddings
• A low alignment score flags potential citation laundering or misrepresentation

The calculated probability that a cited source document logically supports or entails a specific claim made in AI-generated text. This is a core computational primitive for automated fact-checking pipelines.

• Uses transformer-based NLI models fine-tuned on datasets like MultiNLI and FEVER
• Outputs a probability distribution over entailment, contradiction, neutral
• Aggregated across all claims to produce the overall Verifiable Claim Ratio

A metric that quantifies the degree to which a cited claim aligns with established facts stored in a deterministic knowledge graph like Wikidata or a proprietary enterprise graph. This provides a high-precision anchor for verification.

• Queries structured triples (subject-predicate-object) against the graph
• Returns a binary or continuous score based on exact match or semantic proximity
• Serves as a gold-standard reference point immune to LLM hallucination

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim. Corroboration across diverse, authoritative sources dramatically increases confidence in the Verifiable Claim Ratio.

• Implements a quorum-based voting mechanism across sources
• Penalizes claims supported by only a single source or echo-chamber citations
• Requires a Source Diversity Index to ensure genuine independence of corroborating sources

A measure of the completeness and logical validity of the path from an AI's output claim back through its citations to the foundational, verifiable data. A broken or circular chain invalidates the claim for the purposes of the Verifiable Claim Ratio.

• Citation Chaining Protocol recursively traverses references to primary sources
• Detects circular citations where Source A cites Source B which cites Source A
• Flags dead-end citations that lead to inaccessible or deleted content

A statistical range expressing the certainty that a specific claim originates from a given source. This accounts for ambiguities in the AI's source attribution process, such as when a model cites a document but the specific claim cannot be precisely located within it.

• Calculated using conformal prediction techniques
• Narrows with precise Attribution Granularity Level (e.g., passage-level vs. document-level)
• A wide interval reduces the weight of that claim in the overall Verifiable Claim Ratio